When results mislead

How long can you play well and keep losing?

Even with a 55% chance of winning each match, you still have roughly a one-in-four chance of finishing behind after 20 matches.

CUBE BackgammonPublished 18 October 2026Updated 21 October 2026

After several losses, the results alone may not tell you whether your decisions are improving. To see how long results can hide a small advantage, we will give a player a 55% chance of winning each match and calculate how often they still finish with more losses than wins.

A small edge, over 20 matches

Imagine that you have a 55% chance of winning each match. Your opponent has 45%. Keep those chances the same throughout, and make each match independent of the others.

Over 20 matches, you would expect 11 wins and 9 losses. But 11 is an average across many such sessions. An individual evening can end quite differently.

In this model, the chance of finishing with more losses than wins is 24.9%, or roughly one evening in four.

Another 15.9% end level at 10–10. You finish ahead 59.1% of the time. These percentages are rounded, so they do not quite add to 100.

A calculation from the stated model, rather than results collected from players.

Here, an “evening” means a fixed set of 20 matches of the same length. Each whole match counts as one win or loss, whatever its final score. Finishing behind means nine wins or fewer. It does not mean losing 20 in a row.

Give 100 evenings the same starting chance

Change the advantage or the number of matches, then run the experiment. The calculated chances stay fixed for your settings. You get 100 newly simulated evenings each time.

Try 100 evenings

Model: independent matches with a fixed win chance. These settings are assumptions, not ratings.

Finish behind24.9%

Finish level15.9%

Finish ahead59.1%

The same win chance can produce runs of losses, early leads that disappear and recoveries cut short by the end of a session. The simulated player has the same advantage in every case; the different records come from the random draws.

Even 100 matches leave room for disappointment

Keep the 55% win chance and extend the session to 100 matches. The chance of ending behind falls to 13.5%, about one in seven. More results help an advantage show through, but they do not guarantee a winning record over a chosen stretch.

Our simulated player never gets annoyed enough by a loss to start playing differently. Their chance of winning stays at 55% in every match, and they can still finish behind after 100 matches.

The size of the edge matters too. With a 60% win chance in every match, the chance of finishing behind after 20 is 12.8%. After 100, it is 1.7%.

These are comparisons between fixed sets of matches. If you keep extending a session because you are losing, you have changed the question the calculation answers.

The next match does not owe you a win

In our model, your chance of winning the next match stays at 55% whether you have just lost or won the previous five. Those past results do not change the probability of the next one.

The same distinction appears in the consecutive-doubles puzzle: a surprising run behind you does not change an independent event ahead of you.

Real players are more complicated. You might tire, change opponents, switch match lengths or start making rushed decisions. That is why the 55% is an assumption for this experiment, not a claim about your rating or your next game.

What should you take from a losing evening?

A losing record alone cannot tell you whether you played badly, had an unlucky run, or experienced some of both.

Pick a few decisions you were unsure about. Look at what you were trying to achieve and what the alternatives offered. Would you choose the same play again, before knowing the next roll?

We used the same approach in our test on recorded tournament decisions: compare the moves played with the coach’s advice, without treating a win as proof that a recommendation was good.

Inside a match, taking a double can put the whole result on the current position, while passing can leave other games to play. Our cube example follows those two routes from the same board.

If you want a starting point for reviewing your own choices, the five game plans give you a way to describe what your checkers were meant to do.

The broader relationship between skill and short-run results is well established. We consulted and credit Phil Simborg’s “Luck vs. Skill in Backgammon”. The writing, numerical example, interactive experiment and social graphics here were created for CUBE Backgammon.

About the calculation

We use the binomial distribution: a fixed number of independent matches, each with the same assumed probability of a win. A session finishes behind when wins are fewer than half the matches; an even split is reported separately. The headline probability is 24.92893598% for 20 matches with a 55% win chance each. To find the chance of finishing behind, we add up the probabilities of all the losing records. The result does not come from the 100 simulated evenings shown in the experiment.

This is a model of match results, not a simulation of checker play or a measurement of CUBE Backgammon users. It does not convert Elo or performance ratings into win probabilities. Match length, opponents, fatigue and skill are held fixed by assumption. It measures the record at the end of a fixed session, not the longest losing streak, time spent behind, or a guarantee about any particular player. Every match has equal weight.

Keep exploring

From blackjack to bridge: our ranking of game difficulty for computers

Our order is blackjack, English checkers, backgammon, chess, Go, then full partnership bridge. The surprises are what dice do to backgammon’s search, why Go resisted position evaluation, and how hidden cards complicate bridge.

Equity: how a backgammon engine puts a value on a position

Our backgammon engine cannot follow every possible game to its ending. It needs a useful estimate of what each unfinished position is worth, so it can compare the moves that lead there. That estimate is at the centre of its decisions.

Why seeing all four hands can mislead a bridge computer

Building a backgammon engine raises a tempting question: how much can you settle by simulating more games? Bridge, a game my grandfather played brilliantly, shows why a simulation must also respect what the player can know.

How dice helped a computer learn backgammon

The dice that frustrate backgammon players also helped an early neural network learn the game. They carried its practice games into new positions, giving a program playing against itself more varied experience.

How much of backgammon is luck? It depends who is playing

After a game, CUBE Backgammon shows the winner alongside estimates of who played better and whom the dice favoured. Building that analysis forces us to ask what separating luck from skill can actually mean in a single game.

Why blackjack can be calculated without predicting your opponent

When we compare moves in CUBE Backgammon, we need to consider both future dice and an opponent who chooses how to use them. Blackjack keeps the uncertain cards but fixes the dealer’s behaviour, which changes the calculation considerably.

How computers rewrote backgammon’s opening moves

In the 1990s, computers transformed how people played backgammon, a game with centuries of history. Programs that could simulate thousands of games prompted leading players to rethink long-established advice, even about the opening moves.

24 points and six-sided dice: the mathematics of a backgammon board

Some Roman gaming boards had 36 playing spaces, arranged as six words of six letters. By the fifth century, the Byzantine emperor Zeno was playing on a board with the familiar 24 points, though he used three dice.

How long does it take to master backgammon?

Jeremy Bagai thought he had mastered backgammon within a week. Twelve years later, he published his own book correcting mistakes in five influential books he had learned from, written by authors such as Paul Magriel and Bill Robertie.

Emperor Zeno’s bad roll, reconstructed from a poem

A poem records a Byzantine emperor rolling 2, 5 and 6 and ending up with eight exposed checkers. Reconstructing the position shows how the roll forced him to break all three of his pairs.

A backgammon set in the Mary Rose carpenter’s cabin

When archaeologists excavated the Mary Rose, an English warship from Henry VIII’s navy that sank in 1545, they found a backgammon set among the tools in its carpenter’s cabin.

When backgammon players counted doubles with matchsticks

Some players in the backgammon boom of the 1920s and 1930s tracked doubles with ordinary matchsticks. Period guides show why they needed to agree on what each match meant, while a 1930 advertisement offers a more compact solution.

Why some backgammon players call the dice in Persian

Some backgammon players in the Arab world and Turkey call the dice with words inherited from Persian, sometimes mixed with Turkish. A six and a five make a particularly good example.

Saving a gammon: the race to get one checker off

When a comeback looks unlikely, bearing off just one checker can prevent a double or triple loss, but abandoning a good chance to hit can also be expensive.

Why you can’t always choose which die to play first

With a checker on the bar and a 6–1 to play, entering with the one can leave the six unusable. If the other order lets you play both dice, the rules require you to use it.

Why the match score can change your best backgammon move

At 0–0, a chance to win two or three points can be worth taking a small risk. At 6–6 in a match to seven, one point already wins the match. The same board can call for a different move.

Does the app cheat? How CUBE Backgammon generates its dice

For an ordinary computer match, CUBE Backgammon generates the dice sequence from a random starting value before the first roll. After the match, you can reproduce the sequence and check that the recorded rolls match it.

Two equally likely dice patterns, but one wins 35 times out of 36

We do occasionally think about something other than backgammon, as long as it still involves two dice. The puzzle below is about probability rather than backgammon strategy. “Double, double, non-double” and “non-double, double, double” are equally likely in three preselected rolls. Keep rolling until either appears, however, and the second pattern wins with probability 35/36.

We trained CUBE Backgammon to play more like GNU Backgammon, but it chose worse moves

To improve CUBE Backgammon’s move advice, we trained the neural network that evaluates backgammon positions on more examples. It agreed more often with its teaching engine, GNU Backgammon, but chose worse moves in a separate test that played out possible continuations.

How a subtle bug skewed our backgammon engine test

To check whether a development version of CUBE Backgammon chose better moves, we asked Open Sage, another backgammon engine, to evaluate them. On half the positions, our code supplied the wrong player to roll and distorted the comparison.

The cube and the games you haven’t played yet

Taking a double can put the whole match on one position, while passing leaves another route through the games that follow. A small bear-off example shows how those chances compare.

How strong is CUBE Backgammon’s engine?

Our tests against GNU Backgammon provide evidence about CUBE Backgammon’s checker play. Here is what those results show and what we would need to compare the engine with professional players.

167: how to tell who’s ahead in the race

You don’t need to count fifteen checkers one by one. A few familiar groups can tell you how much ground each player has left to cover.

How a backgammon coach fits in your phone

Tap the Hint button to see the coach’s suggested move. Here’s how it chooses that move and finds an answer while you play.

Testing CUBE Backgammon’s coaching advice on 47 tournament matches

We tested a development version of CUBE Backgammon’s coach on recorded tournament play. A second engine found that some suggested corrections were worse than the moves the players had made.

Backgammon strategy: when to attack, block or run

Your opponent makes an anchor, and the checker you wanted to attack is suddenly protected. Follow that change on the board, then explore when to attack, block, run or wait.

Why two exposed checkers can be safer than one

An extra exposed checker looks like an extra problem. Sometimes it gives your opponent fewer rolls that can hurt you.

Why good players seem to roll better dice

Two ways of playing the same 3–2 leave the same pip count, but different opportunities next turn. In our example, one arrangement lets 15 of the 36 rolls make a new point while keeping the existing ones; the other allows only five.

Why you might want your opponent to hit you

A hit can sometimes buy you time to wait for a chance, without giving up the points you need.

The 36 rolls behind every game